Can Jev 1.14 Distinguish Zero-Sugar from Low-Sugar? Genki Forest Packaging Test

Background: E-Commerce Shelf Auditing Challenge

A product package showing both 500 mL and 600 mL on the front raises a data-entry dilemma: which volume should be recorded? Genki Forest's alien lime electrolyte water lists "Net content: 600 mL (500 mL + 100 mL gift)". Humans instantly grasp the total, but software must extract the numbers and recognize the gift qualifier rather than stopping at 500. Similarly, some lime variants carry "0 sugar" while others read "low sugar" — visually nearly identical — so mistaking one for the other would corrupt the catalog.

Jev 1.14 Capabilities and Test Setup

Jev is a decision model that returns structured classifications with probabilities. TypeSafe added image input and a Box type for target localization on 14 October 2026 (both still experimental). The test used six Genki Forest product images from the official site. Each image was asked five fixed multiple-choice questions: category, flavor, sugar label, total net volume, and whether the image proves a six-pack. Closed candidate sets and judging rules were defined (e.g., gift volume counts toward total; choose "undetermined" if illegible).

Classification Results

Across the six images Jev correctly separated "0 sugar" from "low sugar" for the lime electrolyte water, picked 600 mL for the incremental pack, classified white peach jasmine tea as "reduced sugar" and 900 mL, and never mistook a single-bottle shot for a six-pack.

Structured Output and Visibility Probability

Downstream systems can read sugar_label.choice directly without regex extraction. Each prediction includes a visibility probability (0–1) that helps flag glare or occlusion.

Localization with Box Type

Jev 1.14's Box question type accepts a target name and returns proportional bounding-box coordinates (top, bottom, left, right) plus confidence. For the lime electrolyte water's sugar label the model returned:

{
  "type": "box",
  "box": {
    "top": 0.45,
    "bottom": 0.49,
    "left": 0.44,
    "right": 0.55
  },
  "confidence": 0.88
}

Multiplying by image dimensions lets the backend render overlays. The test asked Jev to locate four targets: whole product (A), product name (B), sugar label (C), and net volume (D).

Rendering the raw coordinates highlights the regions for human verification.

Limitations of Localization

Boxes are not always precise: the net-volume box missed the gift-volume note; the red bean coix water was correctly classified "0 sugar" but its box drifted to the upper right of the text. Correct classification does not guarantee accurate localization, and the API returns coordinates even when the target is absent — business logic must not assume existence just because a box appears.

Comparative Evaluation with Clef Flash 9B and OpenAI GPT-6 Luna Decisions

Two reference models were run on the same images and questions: a local Ollama Clef Flash 9B and OpenRouter's OpenAI GPT-6 Luna Decisions (openai/gpt-6-luna-decisions-20261006). Key disagreements against the ground-truth images:

Sparkling water side glare hid the volume. Clef guessed 500 mL with ~0.10 confidence; Jev and OpenAI correctly returned "undetermined".

Red bean coix water's small "0 sugar" text was missed by OpenAI; white peach jasmine tea was misclassified as sparkling water by OpenAI.

Jev showed internal inconsistency: the low-sugar variant got 600 mL (confidence 0.51) yet its volume visibility probability was only 0.10 — a cross-question conflict that downstream rules must catch.

Out of 30 classification judgments, Jev agreed with each of the other two models on 28.

Practical Application and Caveats

Jev's value lies in "structured pre-screening + region-assisted verification": if a merchant enters "0 sugar" but the model says "low sugar", the system blocks and highlights the sugar-label region for manual review; glare or single-bottle-passing-as-case-pack triggers a re-photo request. However, because the features are still experimental, localization offsets exist, and inter-question confidence conflicts occur, production adoption requires blind-testing on proprietary product images and defining explicit human-in-the-loop rules.